Interference optical path for acquiring multiple information by single exposure and application thereof

Abstract

The application belongs to the technical field of optical precision measurement, automation control and artificial intelligence, and particularly relates to an interference optical path for obtaining multiple information in single exposure and application thereof, the interference optical path comprising: rotating a polarization direction of a light source through a half-wave plate to obtain first linearly polarized light; the first linearly polarized light is evenly divided into reference light and test light by a common beam splitter; the reference light passes through a quarter-wave plate back and forth to obtain second linearly polarized light orthogonal to the first linearly polarized light; the test light is decomposed into third linearly polarized light and fourth linearly polarized light by a Wollaston prism; the third linearly polarized light and the fourth linearly polarized light irradiate on a measured object, pass through the Wollaston prism again after reflection, and are combined with the second linearly polarized light at the common beam splitter; the combined light beam is received by a polarization resolution camera and records interference signals.The application can realize absolute phase reconstruction in single exposure, and is suitable for measuring refractive index distribution of dynamic or environment-sensitive samples.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of optical precision measurement, automation control and artificial intelligence, and specifically relates to an interference optical path for acquiring multiple information paths in a single exposure and its application. Background Technology

[0002] Non-contact optical interferometry is a core method for obtaining the three-dimensional morphology of an object's surface or its internal refractive index distribution, and it is widely used in precision manufacturing, biomedical imaging, and materials science. However, traditional interferometry techniques face two fundamental challenges in pursuing higher precision and wider applicability:

[0003] 1. The 2π phase ambiguity problem limits absolute measurement capabilities. Interferometers extract phase information by analyzing changes in light intensity, but the output phase is confined to the interval [-π, π). When the change in the measured physical quantity (such as height or optical path difference) exceeds half the wavelength of the light source (λ / 2), the phase undergoes periodic jumps, making it impossible to directly obtain continuous absolute values. Although dual-wavelength or multi-wavelength interferometry can extend the ambiguity-free measurement range through the principle of wavelength synthesis, it requires multiple independent acquisitions of interferograms at different wavelengths. For dynamic processes (such as droplet deformation or thermal deformation within microfluidic chips) or measurement scenarios extremely sensitive to vibration, this multiple-exposure method introduces severe motion artifacts or environmental noise due to time delays, thus compromising the effectiveness of the measurement.

[0004] 2. Existing AI-assisted solutions lack a physical foundation, making it difficult to guarantee accuracy and robustness. In recent years, deep learning has been used to recover phase from single frames or a small number of interferograms. Some studies have attempted to create virtual phase-shift maps using generative models (such as GANs), while others have used end-to-end networks to directly regress the phase. However, these methods generally suffer from "black box" characteristics, ignoring the strict physical laws governing the generation of interferometric signals (such as orthogonal phase shift relations and light intensity models). This leads to poor model performance outside the training data distribution, weak generalization ability, and difficulty in meeting the stringent requirements of metrology for interpretability and reliability of results.

[0005] Therefore, the industry urgently needs a new measurement paradigm that can both overcome the 2π blur limitation and achieve single transient exposure. This paradigm must capture enough information for absolute phase calculation at the hardware level in one go, and deeply integrate prior physical knowledge at the algorithm level to ensure high accuracy and strong robustness of the results. Summary of the Invention

[0006] To compress the acquisition of multi-view, multi-polarization state information, which traditionally requires multiple exposures, into a single exposure, this invention proposes an interference optical path for acquiring multiple information paths in a single exposure. The interference optical path includes:

[0007] A single-wavelength coherent light source is passed through a half-wave plate to rotate its polarization direction, thus obtaining the first linearly polarized light.

[0008] The first linearly polarized light is evenly split into reference light and test light by a common beam splitter;

[0009] After the reference light passes through a quarter-wave plate, it becomes a second linearly polarized light orthogonal to the first linearly polarized light.

[0010] The test light is decomposed into two beams, a third linearly polarized beam and a fourth linearly polarized beam, by passing through a Wollaston prism. The propagation directions have a small angle between them and their polarization directions are orthogonal. The polarization angle of the reference light is half the sum of the polarization angles of the third and fourth linearly polarized beams.

[0011] The third and fourth linearly polarized light irradiates the object under test, and after reflection, it passes through the Wollaston prism again and is combined with the second linearly polarized light at the ordinary beam splitter.

[0012] The combined beam is received and the interference signal is recorded by a polarization-resolved camera.

[0013] To robustly generate physically accurate virtual interferograms from single-exposure data and precisely calculate the absolute phase, this invention also proposes an application of an interferometric optical path for acquiring multiple information from a single exposure. This path acquires two interferograms with orthogonal polarization directions at a first wavelength and their beat frequency interferogram. Based on the two interferograms with orthogonal polarization directions at the first wavelength, two virtual interferograms with orthogonal polarization directions at a second wavelength are generated. The phase shift order is directly predicted based on the beat frequency interferograms of the two polarization directions at the first wavelength. The normal optical path difference (Piston) is calculated based on the generated two virtual interferograms.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. This invention can realize absolute phase measurement under a single exposure. That is, by designing an innovative interference optical path with dual angles and orthogonal polarization, this invention encodes multi-dimensional physical information in a single transient exposure, providing sufficient initial conditions for solving the absolute phase.

[0016] 2. This invention constructs a physical information embedded AI model, which can robustly generate virtual interferograms that conform to physical laws from single exposure data and accurately calculate the absolute phase. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an interference optical path structure for acquiring multiple information paths in a single exposure according to the present invention;

[0018] Figure 2This is a schematic diagram of the application data processing of an interference optical path for acquiring multiple information paths in a single exposure according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention proposes an interference optical path for acquiring multiple information streams in a single exposure. The interference optical path includes:

[0021] A single-wavelength coherent light source is passed through a half-wave plate to rotate its polarization direction, thus obtaining the first linearly polarized light.

[0022] The first linearly polarized light is evenly split into reference light and test light by a common beam splitter;

[0023] After the reference light passes through a quarter-wave plate, it becomes a second linearly polarized light orthogonal to the first linearly polarized light.

[0024] The test light is decomposed into two beams, a third linearly polarized beam and a fourth linearly polarized beam, by passing through a Wollaston prism. The propagation directions have a small angle between them and their polarization directions are orthogonal. The polarization angle of the reference light is half the sum of the polarization angles of the third and fourth linearly polarized beams.

[0025] The third and fourth linearly polarized light irradiates the object under test, and after reflection, it passes through the Wollaston prism again and is combined with the second linearly polarized light at the ordinary beam splitter.

[0026] The combined beam is received and the interference signal is recorded by a polarization-resolved camera.

[0027] In this embodiment, the core of an interference optical path for acquiring multiple information paths in a single exposure lies in utilizing polarization optical elements to simultaneously acquire three interrelated interference information paths under a single exposure. Its optical path structure is as follows:

[0028] 1. Light source and polarization: A coherent light source with a single wavelength (e.g., 633nm) is used, and the emitted light is linearly polarized.

[0029] 2. Polarization modulation: The beam passes through a half-wave plate (HWP), which rotates its polarization direction to 45°;

[0030] 3. Beam splitting: The 45° linearly polarized light is split into reference light and test light by a common beam splitter (BS);

[0031] 4. Reference arm processing: After the reference light passes through a quarter-wave plate (QWP_ref) round trip, its polarization state becomes 135° linearly polarized light orthogonal to the initial direction;

[0032] 5. Test arm processing: The test light passes through a Wollaston prism and is decomposed into two linearly polarized beams with a small angle between their propagation directions (corresponding to two different viewing angles) and mutually orthogonal polarization directions (e.g., 0° and 90°). These two beams illuminate the object under test, are reflected, pass through the Wollaston prism again, and overlap on the imaging plane.

[0033] 6. Interference and detection: The returning reference light (135° polarization) and two orthogonal test beams (0° and 90° polarization) are combined at BS; finally, the interference signal is received and recorded by a polarization-resolved camera (such as an sCMOS camera equipped with a micro-polarization array).

[0034] Specifically, such as Figure 2 The laser generating the coherent light source, the polarization-modulated half-wave plate, the ordinary beam splitter, the Wollaston prism, and the object under test are all positioned on the same horizontal line. A reference arm is positioned perpendicular to this horizontal line, above the beam splitter. The reference arm consists of a quarter-wave plate (whose fast axis is at 45° to the input linear polarization direction) and a reference mirror. The beam from this branch of the reference arm passes through the quarter-wave plate and is reflected back to the quarter-wave plate by the reference mirror, finally returning to the ordinary beam splitter. The test arm consists of a Wollaston prism. After the incident beam passes through the Wollaston prism, it is split into two beams, resulting in a slight... Linearly polarized light with an angle (corresponding to two different viewing angles) and polarization directions orthogonal to each other (e.g., 0° and 90°) illuminates the object under test. After being reflected back to the Wollaston prism, the two beams overlap on the imaging plane and eventually return to the ordinary beam splitter. An imaging lens and a quarter-wave plate (the fast axis of the quarter-wave plate is at 0° to the input linear polarization direction) are also set at the same perpendicular position as the reference mirror, the quarter-wave plate, and the ordinary beam splitter. The light beams after being combined by the ordinary beam splitter pass through the imaging lens and are received and recorded by a polarization-resolved camera (such as an sCMOS camera equipped with a micro-polarization array).

[0035] A polarization camera can transform the interference wavefront input to the camera into polarization states in four directions: 0°, 45°, 90°, and 135°. This is achieved by designing a micro-polarization array in front of the camera target, containing these four polarization directions. This micro-polarization array is composed of multiple 2×2 pixels (made up of tiny polarizers in four directions), allowing simultaneous acquisition of all four polarization directions. It also functions as a polarization analyzer, ensuring that appropriately polarized light is incident on the camera target. Therefore, it can simultaneously obtain the four polarization states of a single point in the input interference field, and sequentially image each 2×2 adjacent pixel on the camera target. Thus, a RAW image with a mixture of four polarization directions can be obtained through this micro-polarization array. This micro-polarization array contains multiple 2×2 pixels representing the four polarization directions: 0°, 45°, 90°, and 135°. Pixels with the same polarization direction can be uniformly extracted and reconstructed from these multiple 2×2 pixels to form a complete interferogram for each polarization direction. Through this optical path, the system can simultaneously acquire the following three key data streams in a single transient exposure:

[0036] Channel A: All pixels in the RAW image with a 0° polarization direction form an interferogram with a wavelength λ polarization direction of 0°. ;

[0037] Channel B: A wavelength λ interferogram composed of all pixels in the RAW image with a polarization direction of 90°. ;

[0038] Channel C: The beat frequency interferogram is composed of all pixels in the RAW image with a 45° polarization direction. This image represents the difference between the interferograms of Channel A and Channel B. Beat frequency interferogram (moiré fringe pattern). It contains rough information on the 2π phase order (N).

[0039] Based on the aforementioned interference optical path, this invention compresses multi-view, multi-polarization state information that would otherwise require multiple exposures into a single exposure, providing a wealth of physical clues for subsequent AI processing that far exceed those of conventional single-frame images.

[0040] This invention also proposes an application of an interferometric optical path for acquiring multiple information in a single exposure. This path acquires two interferograms with orthogonal polarization directions at a first wavelength and their beat frequency interferograms. Based on the two interferograms with orthogonal polarization directions at the first wavelength, two virtual interferograms with orthogonal polarization directions at a second wavelength are generated. The phase shift order is directly predicted based on the beat frequency interferograms with the two polarization directions at the first wavelength. The normal optical path difference (Piston) is calculated based on the two generated virtual interferograms.

[0041] The purpose of this embodiment is to generate a virtual interferogram for wavelength synthesis calculation from three single-exposure data obtained by an interferometric optical path that acquires multiple information from a single exposure, and finally solve for the absolute phase. This invention constructs a synthesis network for generating the virtual interferogram. This synthesis network includes a large receptive field spatial module, a physical depth embedded module, a contrast-aware channel attention module, and a frequency domain amplitude and phase spectrum learning module, specifically including:

[0042] 1. Input and output of the synthesis network:

[0043] Input to the synthesis network: wavelength The light source, and the three-channel data obtained from a single exposure. As training data, when the synthetic network performs inference, it needs to use two mutually orthogonal interferograms and their corresponding beat frequency maps. ,Right now As input;

[0044] The output of the synthetic network: During training, the synthetic network consists of two outputs: a main branch and secondary prediction branches. The main branch is used to predict the outputs based on two mutually orthogonal interferograms. Output a virtual wavelength interferogram; the secondary prediction branch is used to predict based on two mutually orthogonal interferograms. Beat frequency interferogram Predict 2-step phase order; during the inference phase, only two mutually orthogonal interferograms need to be used as inputs to the main branch, and the main branch outputs the generated virtual wavelength interferogram.

[0045] 2. Structure of the synthetic network:

[0046] To ensure the model can simultaneously generate virtual wavelength phase-shifted interferograms and 2π-order interferograms, this invention designs a main-secondary branch parallel architecture. The main branch can contain multiple channels. Through spatial multi-scale mechanisms, physical law embedding mechanisms, contrast-aware attention mechanisms, and frequency domain learning, the main branch (Generator Backbone) is responsible for extracting high-level features from the input. It encodes data from multiple channels of different dimensions in the same feature space using the same method and fuses the encoded features. The fused features are used to generate 0° and 90° virtual interferograms at λ² wavelength. A typical main-branch structure is given below:

[0047] Channel 1: Spatial Global Reception Channel. The input consists of preprocessed interferograms with polarization directions of 0° and 90°. First, the two interferograms are stitched together along the channel dimension to obtain input features containing dual-angle information. Then, this input feature is downsampled using strided convolution to obtain a downsampled dual-angle spatial feature map, used to reduce computation and extract preliminary fringe features. Subsequently, the downsampled feature map is input into the visual perception module, where a large-size convolution kernel (typically not less than 10×10) or strided convolution is used to further expand the receptive field, resulting in a large receptive field feature map containing local fringe details, global fringe distribution, and overall phase change trends. As an optional implementation, the large receptive field feature map is input into an attention module for enhancement. The attention module learns channel weights through global average pooling, global max pooling, multilayer perceptron, or 1×1 convolution, and generates a spatial attention map through convolution. Larger weights are assigned to regions with clear fringes, distinct edges, and significant phase changes, while smaller weights are assigned to noisy, background, or low-information regions. The attention weights are then multiplied element-wise with the large receptive field feature map to obtain the enhanced dual-angle spatial feature map. If a Transformer structure is used, patch merging and internal downsampling can be further performed after attention enhancement, and positional encoding can be added to preserve the spatial positional information of the interference fringes. Simultaneously, self-attention is used to model the correlation between different patches. Finally, the enhanced dual-angle spatial features are mapped to the same feature dimension as other channels through 1×1 convolution or linear projection to obtain a unified spatial feature vector. The output is a unified spatial encoded feature containing the spatial fringe structure and phase-related information of the dual-angle polarization interferogram.

[0048] Channel 2: Implicit Neural Representation Downsampling Channel. The input consists of preprocessed interferograms with polarization directions of 0° and 90°. First, Fourier transforms are performed on both interferograms to obtain their respective amplitude and phase information. Then, at the same pixel location, the spatial coordinates, the amplitude and phase corresponding to the 0° polarization interferogram, and the amplitude and phase corresponding to the 90° polarization interferogram are combined into a feature vector, which serves as the input to the implicit neural representation. Subsequently, this feature vector is input into a multilayer perceptron or Transformer encoder, where the network learns the mapping relationship from "spatial location and its corresponding amplitude and phase information" to "implicit feature representation." This process is not simply pixel compression, but rather encoding the coordinate, amplitude, and phase information of the two interferograms at the same spatial location into implicit features; the output of the multilayer perceptron or Transformer encoder is the implicit feature corresponding to that location. Finally, the implicit features obtained from different locations are aggregated and mapped to a unified dimension through linear projection to obtain an implicit representation feature vector; the output is a unified implicit coding feature containing information on the continuous structure, frequency domain distribution, and phase change of the two-angle interferogram.

[0049] Channel 3: Contrast-Aware Attention Channel. The input consists of preprocessed 0° and 90° polarization direction interferograms. First, convolutional feature extraction is performed on both interferograms to obtain 0° and 90° feature maps. Simultaneously, the mean, standard deviation, local contrast, and relative rate of change of each image are calculated to characterize stripe brightness distribution, brightness fluctuations, edge intensity, and local texture changes (the calculated results are the contrast statistical features). Then, the contrast statistical features corresponding to 0° and 90° are concatenated along the channel dimension to form a dual-angle contrast description vector, which is input into a multilayer perceptron or a one-dimensional convolutional network to learn the contrast-aware weights. These weights do not directly apply to the original pixels but instead apply to the 0° and 90° convolutional feature maps respectively, enhancing high-contrast stripes, obvious edges, and areas with significant phase changes, while suppressing low-contrast or invalid background areas. Finally, the weighted 0° and 90° feature maps are concatenated or weighted and fused along the channel dimension, and a compact dual-angle contrast feature is obtained through global pooling. Then, linear projection or 1×1 convolution is used to map it to the same feature dimension as other channels to obtain a unified contrast-aware feature vector. The output is a unified contrast-encoded feature containing information on dual-angle interference pattern stripe contrast, edge intensity, and local texture changes.

[0050] Channel 4: Frequency Domain Learning Channel. The input consists of preprocessed interferograms with polarization directions of 0° and 90°. First, Fourier transforms are performed on both interferograms to obtain the amplitude and phase maps of the 0° and 90° interferograms. The amplitude map represents the fringe intensity distribution and frequency energy information, while the phase map represents the fringe position changes and phase modulation information. Then, the 0° amplitude map, 0° phase map, 90° amplitude map, and 90° phase map are fed into convolutional layers, large-size convolutional kernels, or Transformer coding structures for feature extraction, yielding 0° amplitude features, 0° phase features, 90° amplitude features, and 90° phase features. Subsequently, the amplitude and phase features under the same polarization direction are concatenated or weighted and fused to obtain 0° and 90° frequency domain features. Finally, the 0° and 90° frequency domain features are concatenated or weighted and fused along the channel dimension to form a frequency domain feature representation containing dual-angle information. Finally, the fused dual-angle frequency domain features are mapped to the same feature dimension as other channels through 1×1 convolution or linear projection to obtain a unified frequency domain feature vector; the output is a unified frequency domain encoded feature containing information on the frequency energy, amplitude distribution and phase change of the dual-angle polarization interferogram.

[0051] During fusion, the unified feature vectors of the four channels are first concatenated along the channel dimension to obtain a joint feature vector. Then, the concatenated features are compressed and recombined through 1×1 convolution or linear projection, allowing information from different channels to complement each other in the same feature space. Attention fusion can also be further introduced, assigning weights to different channels to enhance features that contribute more to the current fringe structure, phase changes, or frequency domain distribution, while suppressing ineffective or noisy features. Finally, a fused feature vector is obtained, which simultaneously contains the spatial structure, implicit representation, contrast changes, frequency domain amplitude / phase information, and dual-angle phase correlation of the 0° and 90° interferograms.

[0052] Finally, the fused feature vector from the four channels is fed into the decoding and reconstruction structure, which can consist of an upsampling layer, a deconvolution layer, a convolutional layer, or a multilayer perceptron. Specifically, the decoder first takes the fused feature vector as input and gradually expands the spatial size through upsampling or deconvolution to recover a high-resolution feature map consistent with the spatial size of the target interferogram. Then, multiple prediction branches are set after this high-resolution feature map to predict different physical quantities. The background intensity prediction branch can consist of a convolutional layer, an activation function, and a 1×1 convolution, used to output a background intensity map at another wavelength. The modulation prediction branch can use the same convolutional structure to output the modulation plot at another wavelength. The modulation index can be reasonably determined by a non-negative activation function; the phase prediction branch can be composed of a convolutional layer, an activation function, and a 1×1 convolution, used to output the phase in the 0° direction. Phase with 90° direction Therefore, the prediction process can be understood as follows: first, the spatial feature map is recovered from the fused features, and then the recovered spatial feature map is used to regress the background light intensity, modulation degree and two sets of phase information at each pixel position through different prediction heads (each prediction branch only needs to be performed once).

[0053] This decoding and recovery process needs to be optimized during the training phase. During training, the predicted values ​​will be... , respectively with , Substituting into the interferogram light intensity formula:

[0054]

[0055] Reconstruct virtual interferograms at 0° and 90° directions at another wavelength.

[0056] The normal optical path difference Piston is calculated based on the two generated virtual interferograms, specifically including the following steps:

[0057] Step 1: Analyze the two interference patterns with orthogonal polarization directions at the first wavelength, i.e., the interference pattern with the polarization direction at 0° at the first wavelength. Interference pattern with the first wavelength and polarization direction at 90° High-precision first-package phase is calculated based on the standard two-step method. .

[0058] Step 2: Generate two virtual interferograms with orthogonal polarization directions under the second wavelength, i.e., virtual interferograms with a polarization direction of 0° under the second wavelength. Virtual interferogram with polarization direction of 90° at the second wavelength The high-precision second wrap phase was calculated using the standard two-step method. .

[0059] Step 3: Calculate the composite wavelength using the principle of two-wavelength interference. , is represented as:

[0060]

[0061] in, The first wavelength; The second wavelength; This indicates the calculation of absolute value.

[0062] And a rough estimate of the Piston Steps was obtained. , is represented as:

[0063]

[0064] in, Piston represents a rough estimate of the normal optical path difference.

[0065] Step 4: Using the coarse step measurement results, obtain the precise secondary correction N, expressed as:

[0066]

[0067] in, This indicates rounding to the nearest whole number.

[0068] Step 5: Calculate the precise normal optical path difference (Piston), expressed as:

[0069]

[0070] in, This indicates the precise normal optical path difference.

[0071] In this embodiment, when training the synthetic network, the wrapping phase is calculated from the two generated virtual interferograms, and an orthogonal phase constraint loss is applied. Then, the levels obtained from the virtual interferograms are compared with the levels obtained from the secondary prediction branches, and a level consistency loss is applied. Through the joint backpropagation of these three loss functions, the learnable parameters in the decoder, each prediction branch, and the four feature extraction channels at the front end are updated. The specific loss functions include:

[0072] 1. Pixel-level reconstruction loss: This loss term directly affects both the real interferogram and the generated virtual interferogram at the polarization angle corresponding to the second wavelength. By calculating the L1 norm of both at the pixel intensity, it constrains the generated virtual interferogram to approximate the interferogram that might be generated in the real physical world in terms of grayscale distribution, fringe intensity, and local details. Specifically, for the real interferogram with a polarization direction of 0° at the second wavelength... With the generated virtual interferogram And the actual interferogram with a polarization direction of 90° at the second wavelength. With the generated virtual interferogram The corresponding pixel-level reconstruction loss can be written as:

[0073]

[0074] in, Indicates pixel position, This represents the total number of pixels involved in the calculation. and These represent the pixel positions of the generated virtual interferograms with polarization directions of 0° and 90° at the second wavelength. The intensity value at that location, and These represent the pixel positions of the actual interferograms with polarization directions of 0° and 90° at the second wavelength. The intensity value at that location.

[0075] This loss function constrains the real interferogram and the generated virtual interferogram at the pixel level at the second wavelength, ensuring that the generated result approximates the interferogram that could be acquired in the real physical world at the corresponding wavelength and polarization angle in terms of pixel intensity distribution. Since the pixel intensity of the interferogram includes not only the brightness variations of the fringes but also implicit information related to the object's surface topography, optical path difference, phase change, and polarization direction, applying L1 reconstruction constraints to the 0° and 90° polarization directions respectively can maintain the physical realism of the generated virtual interferogram in terms of overall grayscale, local fringe details, and edge variations. This loss function provides direct pixel-level supervision for the virtual interferogram, enabling it to not only satisfy physical constraints in subsequent phase calculations but also approximate the real interferogram in image space, thereby improving the interpretability and reliability of the generated image.

[0076] 2. Orthogonal Phase Constraint Loss: This loss term does not directly affect the generated virtual interferogram pixels, but rather the wrapping phase calculated from the virtual interferogram. Specifically, it first applies to the virtual interferogram with a second wavelength and a polarization direction of 0°. Virtual interferogram with polarization direction of 90° at the second wavelength The second wrap-around phase is generated by calculating using the standard two-step phase shift formula. .

[0077]

[0078] in, This represents the background or average light intensity estimated from two virtual interferograms, used to eliminate the influence of the DC component on the phase calculation. Then, this virtual phase is compared with the reference package phase calculated from the real interferogram or physical model at the second wavelength. The orthogonal phase constraint loss, when constrained, can be written as:

[0079]

[0080] in, Indicates pixel position, This represents the total number of pixels involved in the calculation. This means limiting the phase difference to the range of (−π,π) to avoid phase discontinuities at −π and π.

[0081] This loss term mandates a virtual interferogram with a polarization direction of 0° at the second wavelength. Virtual interferogram with polarization direction of 90° at the second wavelength The second wrap phase calculated using the standard two-step phase shift formula The second package phase with a real or physical reference Maintaining consistency is crucial. Since the 0° and 90° interferograms should physically satisfy an orthogonal phase-shift relationship (i.e., their corresponding phase difference should be π / 2), this loss does not simply constrain the pixel similarity of the generated images, but rather imposes a physical consistency constraint on the virtual interferograms at the level of phase calculation results. This loss ensures that the two generated virtual interferograms satisfy a strict orthogonality relationship, allowing them to be directly and accurately substituted into a mature two-step phase-shift algorithm to calculate the wrapper phase at the second wavelength, without introducing additional computational errors due to phase deviations or insufficient orthogonality between the virtual interferograms.

[0082] 3. Order Consistency Loss: During training, this invention also includes a secondary prediction branch. This branch uses a pre-trained lightweight regression prediction structure, taking the input first wavelength beat frequency interferogram as input, to directly predict the prediction order. Meanwhile, in the main branches, virtual interferograms generated at the second wavelength... and The second package phase is calculated, and combined with step 4 of obtaining the absolute phase, the order inferred from the virtual interferogram is obtained. Prediction level with reasoning levels The regression error between them is taken as the level consistency loss, and the corresponding loss function can be written as:

[0083]

[0084] in, Indicates pixel position, This represents the total number of pixels involved in the level calculation. Indicates the position of the secondary prediction branch. The order obtained directly from the prediction. This represents the order inferred from the virtual interferogram generated by the main branches after phase calculation and absolute phase solution. The above form uses mean squared error as the regression loss to penalize the deviation between the predicted and inferred orders; it can also be replaced with the mean absolute error form based on training stability.

[0085]

[0086] This loss term infers the order by comparing the phase information generated by the main branches. The order obtained by directly predicting the second-order prediction branch based on the first-wavelength beat frequency interferogram This is to ensure consistency between the two in the same physical measurement scenario. Due to the hierarchy The order consistency loss is a key physical quantity in the process of recovering the absolute phase from the wrapped phase. Its value directly affects the results of absolute phase unfolding and height reconstruction. Therefore, if there is a significant deviation between the order inferred from the virtual interferogram generated by the main branch and the order obtained by the order prediction branch, it indicates that the generated phase information is inconsistent with the physical order relationship contained in the input beat frequency interferogram. By introducing this order consistency loss, the order inconsistency between the two branches can be penalized during training. This allows the model to not only focus on the generation quality and phase accuracy of the virtual interferogram, but also maintain the physical consistency between the phase order derived from the virtual graph and the input first wavelength beat frequency information, thereby improving the stability and reliability of the model in the absolute phase acquisition process.

[0087] Total Loss Function: Combining the pixel-level reconstruction loss, orthogonal phase constraint loss, and level consistency loss mentioned above, the total loss function during model training can be defined as a weighted sum of three parts. This total loss function constrains the generated second-wavelength virtual interferogram to approximate the real interferogram in pixel intensity, constrains the wrapping phase calculated from the virtual interferogram to satisfy the 0° and 90° orthogonal phase shift relationship, and ensures that the level inferred from the phase information of the main branch is consistent with the level directly predicted by the secondary prediction branch. The corresponding total loss function can be written as:

[0088]

[0089] in, This represents the pixel-level reconstruction loss, used to constrain the consistency of pixel intensity between the virtual interferograms generated at 0° and 90° polarization directions under the second wavelength and the real interferograms. This represents the orthogonal phase constraint loss, used to constrain the wrapped phase, calculated by the standard two-step phase shift formula from the virtual interferograms of the 0° and 90° polarization directions at the second wavelength, to satisfy the physical orthogonality relationship. This represents the level consistency loss, used to ensure that the predicted levels obtained from the secondary prediction branches are consistent with the levels obtained from the trunk branches based on virtual interferogram inference.

[0090] , , These are the weight coefficients corresponding to pixel-level reconstruction loss, orthogonal phase constraint loss, and level consistency loss, used to balance the contributions of different loss terms during model training. Through this total loss function, the model can simultaneously constrain the generated virtual interferogram from three levels: image pixels, phase orthogonality, and absolute phase level consistency, thereby improving the pixel realism, phase computability, and overall physical consistency of the generated result.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An interference optical path for acquiring multiple information streams in a single exposure, characterized in that, The interference optical path includes: A single-wavelength coherent light source is passed through a half-wave plate to rotate its polarization direction, thus obtaining the first linearly polarized light. The first linearly polarized light is evenly split into reference light and test light by a common beam splitter; After the reference light passes through a quarter-wave plate, it becomes a second linearly polarized light orthogonal to the first linearly polarized light. The test light is decomposed into two beams, a third linearly polarized beam and a fourth linearly polarized beam, by passing through a Wollaston prism. The propagation directions have a small angle between them and their polarization directions are orthogonal. The polarization angle of the reference light is half the sum of the polarization angles of the third and fourth linearly polarized beams. The third and fourth linearly polarized light irradiates the object under test, and after reflection, it passes through the Wollaston prism again and is combined with the second linearly polarized light at the ordinary beam splitter. The combined beam is received and the interference signal is recorded by a polarization-resolved camera.

2. The interference optical path for acquiring multiple information paths in a single exposure according to claim 1, characterized in that, The polarization angle of the first linearly polarized light is 45°, the polarization angle of the second linearly polarized light is 135°, the polarization angle of the third linearly polarized light is 0°, and the polarization angle of the fourth linearly polarized light is 90°.

3. An interference optical path for acquiring multiple information streams in a single exposure, as described in claim 1 or 2, characterized in that, The combined beam is imaged on each 2×2 adjacent pixel on the target surface of the polarization-resolved camera. That is, each 2×2 adjacent pixel records the four polarization states of a point in the interference light field after beam combining.

4. An application of an interference optical path for acquiring multiple information streams in a single exposure, characterized in that, Using the interferometric optical path for acquiring multiple information in a single exposure as described in claim 1, two interferograms with orthogonal polarization directions under a first wavelength and their beat frequency interferogram are obtained. Based on the two interferograms with orthogonal polarization directions under the first wavelength, two virtual interferograms with orthogonal polarization directions under a second wavelength are generated. Based on the generated two virtual interferograms, the normal optical path difference Piston is calculated.

5. The application of the interference optical path for acquiring multiple information in a single exposure according to claim 4, characterized in that, The process of generating a virtual interferogram specifically includes the following steps: Two interferograms with orthogonal polarization directions under the first wavelength are stitched together in the channel dimension and then input into the spatial global sensing channel. The spatial global sensing channel includes a cascaded strided convolution module and a large-size convolution module to obtain spatial coding features. Fourier transforms are performed on two interferograms with orthogonal polarization directions under the first wavelength to obtain phase and amplitude information respectively. The amplitude and phase information of each pixel position are concatenated together with the amplitude and phase information of that position in the two interferograms and input into the two interferograms with orthogonal polarization directions under the first wavelength to obtain implicit coding features. Feature maps are extracted from two interferograms with orthogonal polarization directions under the first wavelength using a convolution module. The mean, standard deviation, local contrast and relative rate of change of the two feature maps are calculated respectively. These features are then concatenated and input into a multilayer perceptron or a one-dimensional convolutional network to obtain contrast perception weights. The two feature maps are then weighted respectively. The weighted feature maps are concatenated in the channel dimension to obtain the contrast encoding features. Features are extracted from the phase and amplitude information of two interferograms with orthogonal polarization directions under the first wavelength. Then, the phase and amplitude information under the same polarization angle are spliced ​​or weighted to obtain the fused features. The fused features under the two polarization angles are then adjusted and spliced ​​or weighted to obtain the frequency domain coding features. Spatial domain encoded features, implicit encoded features, contrast encoded features, and frequency domain encoded features are uniformly mapped to a common feature space of the same dimension through 1×1 convolution, linear projection, or multilayer perceptron. They are then concatenated along the channel dimension to obtain a joint feature vector. The fused feature vector is then fed into the decoding and recovery structure, and through upsampling, deconvolution, or multilayer perceptron, it is gradually restored to a high-resolution feature map with the same spatial size as the target interferogram. The high-resolution feature map is input into three prediction heads to obtain background light intensity information, modulation information, and two mutually orthogonal phase information at another specified wavelength. The background light intensity information and modulation information are combined with the two mutually orthogonal phase information to calculate the interference pattern of two orthogonal polarization directions at another specified wavelength.

6. The application of the interference optical path for acquiring multiple information in a single exposure according to claim 5, characterized in that, The loss function used to train each module during the virtual interferogram generation process includes the L1 norm of the pixel intensity of the virtual interferogram and its real interferogram, the orthogonal phase constraint loss, and the regression loss between the phase shift order calculated based on the two generated virtual interferograms and the phase shift order predicted directly from the beat frequency interferogram.

7. The application of the interference optical path for acquiring multiple information in a single exposure as described in claim 6, characterized in that, The total loss function for training each module during the virtual interferogram generation process is expressed as: in, The total loss function used for training each module in the virtual interferogram generation process; The L1 norm of the virtual interferogram and its real interferogram pixel intensities; The loss function is orthogonal phase constraint, where N is the number of pixels in the training samples. This represents a function that limits the phase difference to the range (−π, π]. This represents the second wrapping phase calculated from the virtual interferogram generated at the second wavelength at the p-th pixel position. This represents the true second wrap-around phase at the second wavelength at the p-th pixel position; This indicates the absolute value; , , These are the weight coefficients corresponding to pixel-level reconstruction loss, orthogonal phase constraint loss, and level consistency loss, respectively.

8. The application of the interference optical path for acquiring multiple information in a single exposure according to claim 5, characterized in that, The process of calculating the absolute phase based on virtual interferometry includes: Calculate the first enclosed phase at the first wavelength based on two interferograms with orthogonal polarization directions at the first wavelength. Calculate the second wrap phase at the second wavelength based on two virtual interferograms with orthogonal polarization directions at the second wavelength. Calculate the combined wavelength of the first wavelength and the second wavelength, and calculate a rough estimate of the normal optical path difference Piston based on the combined wavelength, the first wrapping phase and the second wrapping phase; Based on a rough estimate of the normal optical path difference Piston, the first wavelength, and the first wrapping phase Piston calculates the precise normal optical path difference.

9. The application of the interference optical path for acquiring multiple information in a single exposure as described in claim 8, characterized in that, The rough estimate of the normal optical path difference Piston, calculated based on the synthesized wavelength, the first wrapping phase, and the second wrapping phase, includes: in, This represents a rough estimate of the normal optical path difference, Piston. The synthesized wavelength; This is the first package phase; This is the second wrap-up phase; The first wavelength; The second wavelength; This indicates the calculation of absolute value.

10. The application of an interference optical path for acquiring multiple information paths in a single exposure as described in claim 8 or 9, characterized in that, Based on a rough estimate of the normal optical path difference Piston, the first wavelength, and the first wrapping phase The precise calculation of the normal optical path difference (Piston) includes: in, Indicates the precise normal optical path difference; The first wavelength; For precise secondary correction; This represents a rough estimate of the normal optical path difference, Piston. This indicates rounding to the nearest whole number.

Images